Quorum Table
Split five servers into two groups and see why changes need a majority.
TryWait for a leader, then partition the network. With reduced motion, use Step model first.
Watch forThe three-server group can confirm a change; the two-server group must wait. Heal the network to reconnect them.
Standalone teaching model; it represents no production implementation.
Interactive model
Skill practicedReason about quorum safety under partitions
Transferable practiceIndependent systems-design exercise: explore quorum safety when replicas cannot communicate.
Understand the result
Five servers keep matching copies of a shared history. Break the connection between two groups to see why only a sufficiently large group can confirm the next change. Technical detail: this teaching model demonstrates Raft leader election, log replication, quorum safety, and how a former leader catches up after it learns a higher term.
Controls
- Partition groups
- The model starts with N0 and N1 in Group A, and N2–N4 in Group B. Use the node buttons to change that split before partitioning; assignments lock while the network is partitioned.
- Partition network
- Cuts all communication between Group A and Group B in both directions — enabled once each group has at least one node and the cluster has elected a leader.
- Heal partition
- Restores communication between all nodes so they can reconverge.
- Reseed
- Starts a fresh seeded election and replication run, clears any active partition, and lets you compare a different deterministic scenario.
- Step model
- When reduced motion is enabled, advances the simulation in four paced frames while automatic animation stays paused.
What to watch for
- The starting split is two servers in Group A and three in Group B. Select Partition network: the larger group keeps confirming changes; the smaller group waits until you reconnect them. With reduced motion, use Step model until a leader appears.
- The smaller group cannot confirm a change on its own. In the technical view, a former leader can remain visible until it learns about a newer term.
- Click Heal partition — all 5 nodes converge to the same commit index within a few seconds.
- The Split-brain check reads PASS throughout: no term ever shows two nodes both claiming leader.
Why it happens
This compact model represents messages and elections in coarse ticks and proposes entries automatically. A leader needs votes from a strict majority, and entries need majority replication to commit. A minority cannot commit new entries, though a former leader can retain an uncommitted local suffix until it learns a higher term. Timing and delivery are simplified; this illustrates quorum safety rather than implementing production Raft.
What this exercise helps me practice
This standalone systems-design exercise is not a shipped portfolio system. It shows how I reason about consensus safety under partitions: distinguish what may stay available from what is safe to commit, then trace a single committed history as nodes rejoin.
See the broader engineering brief